SYSTEM ONLINEFRAUD-XAI · CHAOS LABBUILD chaos-lab-fixture-0.1.0

Break the model
before the world does.

Test how a fraud model responds to temporal change before it is released.

Synthetic demonstration · no sign-up required

Temporal Stress Chamber · live preview
Temporal Stress ChamberThree evidence channels, discrimination, calibration, explanation stability, travel from the reference window through a scenario gate into the stressed comparison window. Line thickness and vertical deflection encode the magnitude of change in each channel; a red path shows the injected scenario entering at the gate.scenario gatetransformation boundary · no operator armed yetDiscrimination · PR-AUCbaseline 0.667Δ +0.000stressed 0.667Calibration · ECEbaseline 0.037Δ +0.000stressed 0.037Explanation stability · Attribution-profile cosine similarity (proxy)baseline 1.000Δ +0.000stressed 1.000reference windowstressed comparison window
Table below lists the same values for each evidence channel.

This is the baseline. Apply a scenario to compare the three evidence channels.

SYSTEM · 04 CAPABILITIES

A stress-testing chamber, not a scoreboard.

01

Temporal stress

Age a model against synthetic drift, staleness and channel migration before it meets production data.

02

Blind-spot search

A seeded, deterministic search surfaces scenarios where discrimination and calibration diverge quietly.

03

Separated evidence

Discrimination, calibration and explanation stability stay distinct — never collapsed into one score.

04

Reproducible cases

Every run is addressable by scenario, seed and computation version, so results can be replayed exactly.

PROCESS · 01–03

Test. Compare. Record.

01

Apply a scenario

For selected rows, partially replace the live velocity signal with correlated merchant risk as a fallback, then rerun the logistic model; separately expose a subset of behavioural inputs to lagged observations.

02

Compare the evidence

Review discrimination, calibration and explanation stability separately. The app does not combine them into a single score.

03

Keep a record

Save the scenario, seed and computation version in a reproducible evidence case. This example uses scn-channel-migration-staleness-01, seed 20260903 and chaos-lab-fixture-0.1.0.

Blind-spot frontier
Scatter of 24 deterministic scenario trials from the seeded blind-spot search, plotting absolute discrimination change against calibration deterioration.

24 valid trials shown of 24 run · seed 20260903

Evidence case

Scenario
scn-channel-migration-staleness-01
Seed
20260903
Computation
chaos-lab-fixture-0.1.0

The evidence case records the result, its provenance and its limitations. It does not make an approval decision.

CLASSIFICATION

Scope boundary

This is

  • A controlled scenario / counterfactual stress-testing environment for pre-release and shadow-mode validation.
  • A tool that keeps discrimination, calibration and explanation stability as separate, inspectable evidence.
  • A seeded demonstration built entirely from synthetic fixture data.

This is not

  • A real-time transaction blocking system.
  • A digital twin of any institution, or a forecast of future fraud.
  • A source of composite trust, readiness or approval scores.

Read the full method, limitations, and claim boundaries on the method page.